Counting Crowd with Fully Convolutional Networks

Crowd counting is useful and widely applied in video surveillance. It remains a challenge task for the characteristics of crowd, such as severe occlusions, scene perspective distortions and so on. The existing state-of-the-art methods are spatial information ignored or spatial scale reduced. To solve these problems, we proposed a novel Fully Convolutional Networks (FCN) model, which is end-to-end learned on image patches by regressing crowd density distribution. Our crowd FCN model can output high-precession crowd density map and the crowd quantity can be integrated by the map. Besides, to handle the problem of scene perspective distortions, we proposed an unbiased density ground truth generation method. The experiment results demonstrate that our crowd counting method achieved the best accuracy on the WorldExpo'10 crowd dataset compared with other state-of-the-art methods.